mfakto: an OpenCL program for Mersenne prefactoring

[kracker]

Strange, on James's gpu benchmark page, it says the 7970 is faster than the GTX 580? or bad benchmarks?
GTX 580 is listed at 316, the 7970 is listed at 329 (small diff, but still)

[Bdot]

Originally Posted by kracker

Strange, on James's gpu benchmark page, it says the 7970 is faster than the GTX 580? or bad benchmarks?
GTX 580 is listed at 316, the 7970 is listed at 329 (small diff, but still)

After 3 different owners of cards with the new GCN have their benchmarks at James' site, the AMD GCN data appears about right. I also started discussing this with James, because I think the 7970 should be matching a 570, but not 580. Also, there's some mismatch between the older and the newer AMD cards, which I'll try to provide better data for soon.

Is it possible that most of the nvidia benchmarks are from older versions of mfaktc and not taking into account the latest improvements? I'd suggest to wait for TheJudger releasing the next mfaktc version, and then reminding some nvidia-owners to submit their benchmarks. This should correct the picture, putting 580 and possibly also 570 in front of 7970 again.

[kracker]

Originally Posted by Bdot

After 3 different owners of cards with the new GCN have their benchmarks at James' site, the AMD GCN data appears about right. I also started discussing this with James, because I think the 7970 should be matching a 570, but not 580. Also, there's some mismatch between the older and the newer AMD cards, which I'll try to provide better data for soon.

Is it possible that most of the nvidia benchmarks are from older versions of mfaktc and not taking into account the latest improvements? I'd suggest to wait for TheJudger releasing the next mfaktc version, and then reminding some nvidia-owners to submit their benchmarks. This should correct the picture, putting 580 and possibly also 570 in front of 7970 again.

Ah, I see. How about the GHz edition? it I believe uses about the same power as a 580 on load

[LaurV]

Some of these new AMD thingies are bloody fast... They can't compete yet with nVidia when it come to CudaLucas, due to missing opencl-fft library, but when talking trial factoring, I don't see why they won't be able to keep the pace and even take the lead. I am still advocating for nVidia, that is what I have and love, but talking to people (not only here on the forum) they all tell me that I may be left behind soon... AMD is certainly beating nVidia at price/performance reports (I am not talking about brand new things where the price is given by the "fashion", I am talking about "mature, established" products where the prices stabilized).

[JacksonML]

Please add 32-bit. At the moment I only have a 32-bit computer. Also, would I be able to take the current jobs on Prime95 and move them over to this GPU version? Please answer as that would be helpful

[diep]

Originally Posted by LaurV

Some of these new AMD thingies are bloody fast... They can't compete yet with nVidia when it come to CudaLucas, due to missing opencl-fft library, but when talking trial factoring, I don't see why they won't be able to keep the pace and even take the lead. I am still advocating for nVidia, that is what I have and love, but talking to people (not only here on the forum) they all tell me that I may be left behind soon... AMD is certainly beating nVidia at price/performance reports (I am not talking about brand new things where the price is given by the "fashion", I am talking about "mature, established" products where the prices stabilized).

AMD is horrible in 32x32 bits integer multiplication. It needs 4 units to cooperate, so you lose factor 16 of the single precision floating point throughput they claim, as it is 2 instructions, and each instruction requires 4 units and they brag factor 2 too much as they count fused multiply-add as 2 flops and you cannot use that in factoring.

The Nvidia Fermi series on other hand loses only factor 4 as it is 2 instructions and they also brag (like everyone) factor 2 because of FMA. Factor 4 advantage for Nvidia.

A 7970 gpu is clocked 1Ghz and has 2048 cores. So factor 4 reduction it is similar to a 512 core Fermi.

The 512 core fermi's are dirt cheap on ebay i would guess. Let me look it up.

You can get them for $350 on ebay, but if you bid $250 i bet you might get one as well.
That's for a geforce GTX 580

p.s. note that on paper AMD cards are fast, yet that's based upon statements from AMD and they didn't deliver.
For example they brag that they have a topbits instruction in hardware, and indeed it exists, yet OpenCL doesn't
support this instruction. So opencl misses a mul24_hi() built in function that casts onto the equivalent instruction of
the GPU.

If this would exist AND they didn't lie about its speed (someone who doesn't deliver usually has something to hide - like that
it's not usable practical), then it would allow 24x24 bits == 48 bits.

It's possible to make something fast out of that.

YET THIS IS ALL PAPER THEORIES SO FAR.

AMD just didn't deliver and they are nearly bankrupt. That's the sad truth.
Happens if you get taken over by Arabs.

All codenames are arab names nowadays at AMD.

Latest CPU called after their biggest investor: Abu Dhabi...

I supose however that someone would be interested in taking over ATI after AMD goes bankrupt within a few months.
If that's not another Arab dude, then there is still hope for ATI to improve its helpdesk and services to those who
like to do gpgpu at it.

I will not quote communication i had with AMD past years here right now, yet if i would do it - you'd never again buy AMD as a homeuser who wants to work stable and/or do gpgpu.

The 6970 i bought just after it released, that was the last AMD card i bought.
It took months until there was a driver. When the driver was there, it still didn't work at all OS-es.

For example windows server was not supported. Fine with me - i installed linux.

Debugging in linux not possible very well, as all their tools and toys just work for windows vista or newer.

I have windows vista ultimate installed now at a machine with the AMD card inside - guess what - the machine crashes
regurarly. I had initially every day a bluescreen *after* i put in the AMD card AND installed the AMD drivers (catalyst).

Several times a day a bluescreen - that's more than i can endure.

It improved now - but that's not thanks to AMD, you won't get that at work at your machine like this - i'm sure of it.
A bluescreen is what you'll get! 3 times a day!

How can i 'stable' Trial Factor at such gpu?
As the Tesla's here... ...they're intended for Wagstaff VRB-Reix testing... ...when they idle.

[flashjh]

What is the fastest AMD card anyone is using? In curious about about actual GHz days throughput?

[diep]

Originally Posted by flashjh

What is the fastest AMD card anyone is using? In curious about about actual GHz days throughput?

I've got a HD6970 here. It's 1536 cores and i can manually set the clockspeed of the gpu. Set at 500Mhz now. Maximum 880Mhz at production level, yet i can overclock it when needed (never in your life do that with gpgpu i'd say).

The HD7970 series, you see that's another problem with AMD. their technical manuals get editted by the most junior girl the Indian helpdesk just employed for 75 dollar a month (god i wish i was joking - it's the literally truth - one of the guys there complained to me that he was a senior and making $150 a month).

So from the documentation, that's basically a rewrite from some old ATI document dating back to 2007. All kind of diagrams the junior girl didn't understand have been stripped out of the documentation, so the documentation is total gibberish now with no hard clear description of how a GPU functions, simply as the junior helpdesk girl didn't understand this herself.

Over the past few years they regurarly renamed things. Streamcores were renamed to PE's (processing elements) and nowadays are again called in latest incarnation document again cores. A compute core in fact is 4 cores and a compute unit in previous document was 64 PE's. The latest document again mixes that up and cross mixes the word compute core where they used PE previously.

Obviously you learn nothing new from all this.

So basically see the 7970 as the same gpu like the 6970, just the 7970 has 2048 cores and is a tad higher clocked thanks to moving to a new proces technology.

In itself they could have easily put 3000 cores on it if they wanted to, clocked at say 800Mhz, yet they wanted 1Ghz+, so that's why the increase in number of cores is so little.

So you can simply lineair extrapolate results. There is simply no documentation refuting that isn't the case :)

[kracker]

Really? have you even SEEN benchmarks or know what they are?

@flashjh This will help, although most of the nvidia cards are not mfaktc 0.19 I think, but just to see how many GHZ/days you can get from them.

[diep]

Originally Posted by kracker

Really? have you even SEEN benchmarks or know what they are?

@flashjh This will help, although most of the nvidia cards are not mfaktc 0.19 I think, but just to see how many GHZ/days you can get from them.

You realize some of my benchmarks made it into spec?

[kracker]

Originally Posted by diep

You realize some of my benchmarks made it into spec?

I didn't understand that.

But a 6970 and a 7970 is *not* almost same if you look at those benchmarks.